Baseline validation of a bias-mitigated loan screening model based on the European Banking Authority's trust elements of Big Data & Advanced Analytics applications using Artificial Intelligence

TL;DR

This study validates a bias-mitigated AI-based loan screening model using minimal, non-discriminatory features, demonstrating high accuracy in detecting bad loans while aligning with EU trust principles.

Contribution

It introduces a novel loan screening model that uses minimal, non-discriminatory features and complies with EU trustworthy AI principles, with baseline validation results.

Findings

High sensitivity (0.91) in detecting bad loans

Strong specificity (0.90) in classifying exposures

Supports bias mitigation and consumer protection in credit risk assessment

Abstract

The goal of our 4-phase research project was to test if a machine-learning-based loan screening application (5D) could detect bad loans subject to the following constraints: a) utilize a minimal-optimal number of features unrelated to the credit history, gender, race or ethnicity of the borrower (BiMOPT features); b) comply with the European Banking Authority and EU Commission principles on trustworthy Artificial Intelligence (AI). All datasets have been anonymized and pseudoanonymized. In Phase 0 we selected a subset of 10 BiMOPT features out of a total of 84 features; in Phase I we trained 5D to detect bad loans in a historical dataset extracted from a mandatory report to the Bank of Italy consisting of 7,289 non-performing loans (NPLs) closed in the period 2010-2021; in Phase II we assessed the baseline performance of 5D on a distinct validation dataset consisting of an active portolio of 63,763 outstanding loans (performing and non-performing) for a total financed value of over EUR 11.5 billion as of December 31, 2021; in Phase III we will monitor the baseline performance for a period of 5 years (2023-27) to assess the prospective real-world bias-mitigation and performance of the 5D system and its utility in credit and fintech institutions. At baseline, 5D correctly detected 1,461 bad loans out of a total of 1,613 (Sensitivity = 0.91, Prevalence = 0.0253;, Positive Predictive Value = 0.19), and correctly classified 55,866 out of the other 62,150 exposures (Specificity = 0.90, Negative Predictive Value = 0.997). Our preliminary results support the hypothesis that Big Data & Advanced Analytics applications based on AI can mitigate bias and improve consumer protection in the loan screening process without compromising the efficacy of the credit risk assessment. Further validation is required to assess the prospective performance and utility of 5D in credit and fintech institutions.